DIVISION S-5—PEDOLOGY

Differentiating Soil Types Using Electromagnetic Conductivity and Crop Yield Maps

C. M. Anderson-Cook,* M. M. Alley, J. K. F. Roygard, R. Khosla, R. B. Noble, and J. A. Doolittle

ABSTRACT soil maps. Robert (1993) discusses the viability and cost-effectiveness of a number of options available for creat-Variable rate technology enables management of individual soiling these maps. In the mid-Atlantic coastal plain, soiltypes within fields. However, correct classification of soil types for

mid-Atlantic coastal plain soils are currently impractically expensive property changes within fields are often abrupt becauseusing an Order I Soil Survey, yet variable rate fertilizer application of the alluvial nature of the soils. Fields typically containbased on soil type can be highly effective. The objectives of this study two or more soil types. These soil types vary in produc-were to determine if apparent electromagnetic conductivity (ECa) tivity potential because of differences in soil properties,alone or combined with previous year crop yields using global position- particularly those properties that influence soil water-ing system technology can provide a useful alternative to detailed soil holding capacity, such as clay content. County soil sur-mapping. The site contained alluvial soils ranging from Bojac 1 and 2

veys often fall short of the spatial accuracy required to(coarse-loamy, mixed, thermic, Hapludults) to Wickham 3 and 4 (fine-realize the benefits of variable rate technology, howeverloamy, mixed, thermic, Ultic Hapludalfs). The two fields totaled ap-these surveys provide excellent information regardingproximately 24 ha. A statistical nonparametric classification method,the soil types that may be encountered in the field.called recursive binary classification trees, was used to determine

how well soil types could be classified. Electromagnetic conductivity Order 1 soil surveys (Soil Survey Staff, 1993) providereadings and crop yields were positively correlated. Broad patterns accurate soil maps, however they are generally expen-in the relationship between soil types and ECa readings and crop yields sive to make, or difficult to obtain (Robert, 1993).existed for all crop combinations considered. Lower ECa readings and Variable rate fertilizer applications in the mid-Atlan-crop yields corresponded to the Bojac soils, while higher ECa readings tic region are generally based on soil test values for P,and crop yields were categorized as Wickham soils. Electromagnetic K, and lime applications, and soil yield potential for Ninduction alone correctly classified the soils into broad categories of

fertilizers. Soil sampling strategies for P, K, and limeBojac or Wickham with over 85% accuracy. When ECa was combinedapplications range from grid sampling to managementwith crop yield data, correct classification rose to over 90%. Morezone sampling strategies that are based on soil types.precise classification into Bojac 1, Bojac 2, and Wickham soils yieldedThese sampling techniques may then be used to deter-slightly lower correct classifications ranging from 62.6 to 81.2% for

ECa alone, and 80.3 to 91.5% when combined with various crop yields. mine fertilizer application rates ranging from one rateto a different rate for each sampled location. Recentwork has shown that in cases with relatively little system-atic variation across a field, a composite-by-soil samplingVariable rate fertilization strategies based onapproach can provide effective fertilizer recommenda-soil type have potential to lower fertilizer inputstions with lower sample numbers (Anderson-Cook etand increase profits in the mid-Atlantic coastal plain.al., 1999). The composite-by-soil approach involves col-Conventionally, fertilizer application to a farm-field islecting multiple samples from each known soil type andbased upon the most productive soil in that field. Conse-using a composite sample by soil type as an aggregatequently, low productivity areas of the field may be overmeasure of soil fertility and corresponding fertilizer rec-fertilized. The use of variable-rate fertilizer applicatorsommendation for each soil type within a field. The com-equipped with global positioning systems (GPS) andposite by soil type sampling approach has reduced sam-georeferenced soil maps enables nutrient applicationspling and analytical costs, compared with grid samplingthat better match crop requirements on various soilsstrategies, but accurate soil maps are required to employwithin a field. Applying fertilizer based on yield poten-this strategy (Anderson-Cook, 1999).tials associated with specific soil factors such as water-

In this paper we consider the use of an electromag-holding capacity, optimizes yield response to fertilizernetic induction meter, EM38 (Geonics Limited1, Missis-and reduces potential loss to ground and surface waters.sauga, ON, Canada) to develop soil maps for use inSoil specific fertilization strategies require accuratevariable rate fertilizer application. McNeill (1986) hasdescribed principles of operation for the EM38. Appar-

C.M. Anderson-Cook, Dep. of Statistics (0439), Virginia Tech, Blacks-ent electromagnetic conductivity uses electromagneticburg, VA 24061; M.M. Alley and J.K.F. Roygard, Dep. of Crop and

Soil Environmental Sciences (0403), Virginia Tech, Blacksburg, VA24061; R. Khosla, Dep. of Soil and Crop Science, Colorado State

1 Mention of trade names is solely to provide information for theUniv., Fort Collins, CO 80523; R.B. Noble, Dep. of Mathematics andreader and does not constitute endorsement of the instrument orStatistics, University of Miami (Ohio) Oxford, OH 45056; and J.A.product by Virginia Polytechnic Institute and State University, Colo-Doolittle, USDA-Forest Service, 11 Campus Blvd. (Suite 200), New-rado State University or USDA.ton Square, PA 19073-3200. Received 1 Mar. 2001. *Corresponding

author (candcook@vt.edu).Abbreviations: ECa, apparent electromagnetic Conductivity; GPS,global positioning system.Published in Soil Sci. Soc. Am. J. 66:1562–1570 (2002).

1562

ANDERSON-COOK ET AL.: DIFFERENTIATING SOIL USING CONDUCTIVITY AND YIELDS 1563

prediction is possible for new values of the explanatory vari-energy to measure the apparent conductivity of a col-ables. However, in this case the response is not a continuousumn of soil material to a specific observational depthmeasurement, but rather a category. The classification trees(0.75–1.5 m). Electromagnetic conductivity measurementsare an automated statistical tool for partitioning the rangeshave been used as a surrogate measure for soil proper-of explanatory variables, X1, ···, Xk, into a unique classificationties such as salinity, soil moisture content, topsoil depth, of the response, in this case soil type. Based on the classifica-

and clay content (Sudduth et al., 2001). Scanlon et al. tion suggested by the tree, each combination of values for the(1999) showed that EM38 measurements correlate well variables, x1, ···, xk, suggests a single category for the response.with clay content and soil moisture content, two proper- As with a regression model, the response is required for theties that can vary between soil types and influence crop model fitting stage to determine the nature of the relationship

between response and explanatory variables. When the modelyield potential of soils. Soil salinity as a factor influenc-is used for predictions, only the explanatory variables, x1, ···,ing EM38 measurements is generally minimal in thexk, are required.mid-Atlantic coastal plain because of regular leaching

The tree is constructed from the observed data with a re-events in this high rainfall region (�1000 mm annually).cursive binary partitioning algorithm (Breiman, 1984; ClarkZalasiewica et al. (1985) used ground conductivity toand Pregibon, 1992). The algorithm works to repeatedly splitimprove geological mapping, and Rhoades (1981) dis- the data into advantageous groupings to optimize the predic-

cussed predicting soil electrical conductivity from soil tion ability of the explanatory variables to describe the re-type. In addition, Rhoades et al. (1989) produced a sponse. At each stage of the algorithm, the split of the rangesmodel describing the relationship between bulk soil of the X1, ···, Xk’s into an additional classification region iselectrical conductivity and electrical conductivity of soil based on maximizing the improvement in the correct classifica-

tion rate for the response. To ease the computational demandswater. Doolittle et al. (1995) evaluated the use of ECaof tree construction and to make the interpretation of resultsmeasurements with GPS to conduct reconnaissance soilmore straightforward, each split of the data involves only amapping. Thus, EM38 measurements have potential forsingle true-false decision involving a single variable. Splittingmapping soil zones of differing productivity potentialthe ranges of the explanatory variables continues in the algo-for variable rate fertilization strategies.rithm until subsequent splits fail to yield sufficient improve-In this research, we compare the effectiveness of the ment in the model to justify continuing.

ECa readings to correctly differentiate soil types relative The result of the classification tree is a series of decisions,to an Order I Soil Survey (Soil Survey Staff, 1993). We displayed in an inverted tree. Figure 1a gives a sample treealso investigate combining ECa readings with previous using two explanatory variables, X1 and X2. Starting at thecrop yield maps obtained using a combine yield monitor top of the figure or root, we determine whether to go left for

a true decision or right for a false decision at each branch,and GPS technology that may further improve soil cate-until we have reached the bottom leaves, which tell us thegorization for use in precision agriculture.classification for each particular combination of values for X1

and X2. For example, if x1 � 25 and x2 � 22, we would predictTHEORY the response as Category 1, while if x1 � 33 and x2 � 14, we

would predict Category 2.The goal of utilizing the ECa meter and GPS is to provideAn alternate summary of the tree is given in Fig. 1b wherea mechanism for accurately predicting soil type in a time and

the ranges of X1 and X2 values observed in the data are parti-cost-effective manner. The EM38 induction meter has beentioned into regions. Note that each horizontal or vertical lineused to measure topsoil depth in claypan soils in Missouricorresponds to a single decision involving either X1 or X2,(Doolittle et al., 1994) as well as the depths of sand depositionrespectively in Fig. 1a. To use this summary to predict, wefollowing flooding (Kitchen et al., 1996). Jaynes et al. (1995)would locate a combination, say x1 � 25 and x2 � 22, on theand Sudduth et al. (1999) investigated the relationship be-plot and then predict the category based on the label for thattween yields and topsoil depth as estimated by electromagneticregion; that is, Category 1.induction methods. The Iowa study of Jaynes et al. (1995)

The major advantage of using classification trees to catego-showed yields of corn (Zea mays L.) and soybean (Glycinerize the responses is the flexibility of this nonparametricMax. L.) were correlated with electromagnetic measurements,method to accommodate many different functional forms ofbut correlations were not consistent from year to year. Thea relationship and avoid imposing a prespecified structureauthors indicated that the lack of consistency of the correla-on the data. Unlike regression (for continuous responses) ortions was due to variable spring moisture conditions associatedlogistic regression (for categorical responses), no continuity,with poorly drained soil conditions. Sudduth et al. (1999) re-smoothness, or other global structure for the relationship isported that corn grain yields were correlated to EM38 mea-assumed. Additionally, the method is easily implemented us-surements (r � 0.93), especially in years where water availabil-ing S-Plus software by Mathsoft, (S-Plus, 2000), with existingity was a limiting factor.functions to construct and display the trees.In this study the results of an Order I Soil Survey were used

to define soil type known to require different managementstrategies to optimize crop yields. We sought to develop a MATERIALS AND METHODSmodel to compare the ECa readings to the Order I Soil Survey.Statistical methods were utilized to determine how precisely The soils of the experimental site (Fig. 2) in the Virginia

coastal plain were determined by an Order 1 soil survey (Nich-the ECa meter responses could classify known soil types. Sincelittle was known initially about the relationship between elec- olson et al., 1998) using the methods of the National Coopera-

tive Soil Survey soil survey manual (Soil Survey Staff, 1993).tromagnetic induction readings and soil type, a nonparametricapproach for modeling the relationship with little imposed The site contained alluvial soils ranging in productivity. The

Bojac 1 soil has higher productivity than the Bojac 2 soil, andstructure was taken.Classification trees are similar to regression modeling, in the Wickham 3 and 4 being the most productive soils in the

study. Soil profile characteristics are given in Table 1. Rangesthat the original data suggests an estimated model from which

1564 SOIL SCI. SOC. AM. J., VOL. 66, SEPTEMBER–OCTOBER 2002

Fig. 1. Summaries of classification tree categorization. (a) Inverted “tree” structure for classifying observations with two explanatory variables,X1 and X2. (b) Partition of the observation space of X1 and X2 into classification regions. Values of X1 and X2 are hypothetical.

in productivity and fertilizer needs are associated with avail- 1980b). The instrument was operated in this study in the hori-zontal mode, which provides an effective measurement depthable water-holding capacity and soil texture (Simpson et al.,

1993). The total study area is approximately 24 ha. of approximately 0.75 m (McNeill, 1992). The ECa meter read-ings were taken on a grid at approximate 30.4-m intervals alongElectromagnetic measurements were made using an EM38

induction meter (Geonics Limited, ON, Canada) on 2 Dec. 21 strips each 18.5 m apart. Global positioning system receiverscoupled to a microcomputer recorded the location of the mea-1998. McNeill (1980a) described the principles underlying the

use of electromagnetic techniques to detect differences in soil surements as well as the measured and temperature-correctedapparent conductivity. The 21 strips are part of a larger studyproperties. The EM38 instrument resembles a carpenter’s level,

is approximately 1 m long, and includes calibration controls and involving multiple crop rotations. Hence, measurements weredivided into the four distinct crops (corn, barley [Hordeuma digital readout of ECa. All ECa measurements reported in

this study were standardized to an equivalent electrical con- vulgare L.], wheat [Triticum aestivum L.], and soybean). TheOrder I Soil Survey results were matched with the ECa read-ductivity at 25 �C as conductivity changes 2.2% �C�1. (McNeill,

ANDERSON-COOK ET AL.: DIFFERENTIATING SOIL USING CONDUCTIVITY AND YIELDS 1565

Table 1. Soil horizon, depth, color, and texture for the four soilsat the experiment site in the northern Virginia coastal plain.

Soil Horizon Depth Munsell color Texture

cmBojac 1 Ap 0–33 10YR 3/4 Loamy sand

Bt1 33–81 75YR 4/6 Sandy loamBt2 81–142 5YR 4/6 Sandy clay loamBC 142–165 7YR 5/8 Loamy sandC 165–203 75YR 6/8 Loamy sand

Bojac 2 Ap 0–30 75YR 3/4 Loamy sandBt1 30–104 75YR 4/6 Sandy loamBt2 104–178 75YR 4/6 Loamy sandC 178–203 75YR 5/6 Loamy sand

Wickham 3 Ap 0–25 75YR 4/4 Sandy loamBt1 225–64 5YR 4/6 Clay loamBt2 64–89 5YR 4/6 LoamBt3 89–140 75YR 4/6 Sandy clay loamC 140–203 75YR 4/6 Loamy fine sand

Wickham 4 Ap 0–33 75YR 3/3 Sandy loamFig. 2. Order 1 soil survey map of the experimental site in the northernBt1 33–61 75YR 4/6 Sandy clay loamVirginia coastal Plain (Nicholson et al., 1998).Bt2 61–89 75YR 4/6 Clay loamBt3 89–140 75YR 4/6 Clay loamings using the GPS locations. Thus the classification of each Bt4 140–178 5YR 4/6 Loam

rectangle area in the output graphs of the cluster analysis to C 178–203 75YR 4/6 Sandy loama soil type was determined by examining the data within thatregion and selecting the soil type which comprised the majorityof observations using the associated GPS location and Order 1 Division of the data into crops separated the 472soil survey data. Since the Wickham 3 and 4 soil types did not observations into four crop groups depending on whichrequire separate management strategies for fertilizer applica- crop yield was considered. Therefore some of the ECation because they have similar water-holding capacities, they observations were used more than once in associationwere combined into a single type. Additionally, the task of with different crops in a rotation. Comparison of ECatrying to separate the Bojac and Wickham soils using ECa and measurements to nearest crop yield measurements forcrop yields was also considered to distinguish the soil types

a given crop and the average of the three nearest yieldsinto broader, easily discernable categories.resulted in classification tree models that were virtuallyBefore any analyses were considered, locations at or withinidentical for both approaches. The data presented here10 m of the boundary between soil types were removed fromrelated ECa measurements to nearest crop yield mea-the data set. As these soil measurements would likely be a grad-

uation between the soil types being studied, they could not surement.be defined as a specific soil type. Thus we were unable to The effectiveness of classification trees using the ECarelate ECa for these specific points to a reference soil type data and crop yields to separate the broad classes offrom the Order 1 soil survey. Therefore, of the initial 524 Bojac and Wickham soils are summarized in Table 3.measurements, only 472 were considered. The percentage of correct classification reported indi-

Yields for a variety of crops during 1998-1999 were deter- cates successful soil type categorization using the classi-mined by harvesting with a John Deere 9610 combine (Deerefication tree approach and how well gross differencesand Company, Moline, IL) equipped with the Greenstar yieldbetween the soil types can be identified using ECa andmonitor (Deere and Company, Moline, IL) and GPS receivercrop yields. For each of the crop systems, either ECa orwith differential correction. Yields and the associated loca-crop yields alone separate the Bojac and Wickham soils.tions were recorded at 1-s intervals by translating flow and

area harvested into an instantaneous yield measure. For each The ECa results are expected because the means of thecrop, a global adjustment for standard payable moisture was ECa measurements within the two soil series are quiteapplied. Since crop yield and ECa data were to be considered different relative to their standard deviations (Table 2).together as possible explanatory variables in a model to predict It should be noted that there is considerable consistencysoil type, GPS locations associated with both variables were in the predictive power of ECa values across the differentused to combine the data. Both the nearest crop yield measure- crops. For all of the crops, between 85 and 95% of thements for a given crop and the average of the three nearest

data are correctly classified using only ECa measure-yields to the measurement locations of ECa were considered.ments. However, in four of the five crops considered,Five separate analyses were considered using the different cropcombining ECa with crop yields gives improvement inrotations. Results are presented for 1998 and 1999 full-season

corn, 1999 barley, 1999 wheat, and 1999 double-crop soybeans.Table 2. Electromagnetic induction measurements by soil type.

Data were temperature corrected to 25 �C.RESULTS AND DISCUSSIONSoil Type

Initial examination of the data showed differences inBojac 1 Bojac 2 Wickham 3 Wickham 4ECa values for the broad classes of Bojac and Wickham

No. observations 80 180 113 99soils (Table 2). The Bojac soils have lower conductivityAverage, mS m�1 2.60 2.84 5.29 4.21associated with lower moisture retention than the Wick-S.D., mS m�1 0.94 0.97 1.23 0.82

ham soils. Differences within the Bojac 1 and 2, and theAverage, mS m�1 2.77 4.79Wickham 3 and 4 soils are less, as would be expected,S.D., mS m�1 0.96 1.19because surface soil textures are similar (Table 1).

1566 SOIL SCI. SOC. AM. J., VOL. 66, SEPTEMBER–OCTOBER 2002

Table 3. Percentage correct classification rate of classification trees for Bojac versus Wickham soil types using apparent electromagneticconductivity (ECa) and crop yields alone and combined.

Crop Number of observations ECa Alone Crop yield alone ECa & crop yield

% correctly classifiedNo-till full-season corn 1998 129 85.3 88.4 91.5No-till full-season corn 1999 197 91.4 88.8 96.4No-till barley 1999 67 95.5 86.6 95.5No-till wheat 1999 211 87.2 83.0 93.4No-till double-crop soybean 1999 259 87.6 70.0 90.3

classification success (Table 3). Yield data generally im- Wickham soil, while the unshaded region predicts theBojac soil. Lines inside each region indicate a separateprove the classification over ECa data alone because

crop yields tend to reflect soil moisture holding capacit- decision in the classification tree, however, we are pri-marily interested in how the broad patterns defined byies for these coastal plain soils. The exception to im-

provement in classification success using yield data was different combinations of ECa readings and crop yieldspredict soil types.the Barley misclassification rate, which remains the

same when using ECa alone or in combination with crop The ECa and yield measurements are significantly andpositively correlated (ranging from 0.16 for wheat toyield data. A misclassification rate below 10% for all

of the crops suggests the effectiveness of this approach 0.54 for corn). This positive correlation is related to theECa relationship with clay content and soil water contentfor distinguishing between Bojac and Wickham soils,

regardless of the crop being grown. at the time of measurement and crop yields relationshipwith clay content and soil water holding capacity. TheFigures 3 and 4 show the partitioning plot for the

classification trees for 1999 no-till full-season corn and values of the ECa increase with increasing clay contentand soil moisture. At the time of ECa measurement in1999 no-till wheat. The x-axis shows the ECa values, while

the y-axis gives the crop yield for the corresponding December the soil water content would have closelyreflected the differences in water-holding capacity be-GPS location. The points on the plot are the data collec-

tion sites corresponding to a particular combination of cause of a recharge of the soil profile by rainfall. Forthis reason the data is not uniformly spread throughoutECa with corresponding crop yield. The relatively high

variability of the yields within soil types is a reflection the ranges of the explanatory variables, as seen by therelative lack of observations in the top-left and bottom-of the 1-s interval readings of the yield monitor being

translated into a kilogram per hectare summary. The right corners of the plots. The combination of low ECa

and low crop yield is consistently categorized as Bojacrectangles form boundary lines around the different soiltype classifications. The shaded region corresponds to soil, while observed data associated with higher ECa

Fig. 3. Classification partition for Bojac versus Wickham soil types using apparent electromagnetic conductivity (ECa) and 1999 no-till full-season corn yields. The classification of each rectangle is determined by whichever soil type comprises the majority of observations.

ANDERSON-COOK ET AL.: DIFFERENTIATING SOIL USING CONDUCTIVITY AND YIELDS 1567

Fig. 4. Classification partition for Bojac versus Wickham soil types using apparent electromagnetic conductivity (ECa) and 1999 no-till wheatyields The classification of each rectangle is determined by whichever soil type comprises the majority of observations.

and crop yield is categorized as Wickham. This was the The 1999 barley classification tree (data not shown)separates the soils with one vertical line, indicating thatgeneral pattern for the separation of Bojac and Wick-

ham soils for all crops evaluated. only ECa is important for soil type delineation for thiscrop. In this way the classification tree can be consideredBecause of the positive correlation between ECa read-

ings and crop yields, high ECa reading corresponding a variable selection technique as well, since it choosesat each stage not to include the crop yields as an informa-to low yields, or a low ECa reading corresponding to high

crop yields were rare. For these combinations, there are tive variable for separating the soil types. Since cropyields are not used in the barley analysis, we have thesome small differences in the classifications for the 1999

corn data (Fig. 3) and the 1999 wheat data (Fig. 4). same correct classification rate for both ECa alone andECa with crop yield trees (as shown in Table 3).The number of branches or decisions in the trees,

denoted by the number of horizontal and vertical lines Table 4 shows how the soil types are classified usingECa and crop yield for both the 1998 and 1999 no-tillon the partitioning plot, indicate the complexity of the

relationship between the explanatory variables and the full-season corn yields. Comparison of results betweenthe corn years shows the consistency of the results fromsoil type response. Since the classification tree algorithm

cannot assign diagonal lines to separate data, it must year to year, and in different locations in the total studyarea. In both years, over 90% of soil types are correctlyuse multiple horizontal and vertical lines for this pur-

pose. The classification tree approach is strongly empiri- classified, with the Bojac soils having a smaller misclassi-fication rate. To obtain the total correct classificationcally based and does not force global structure on the

relationship between explanatory variables and the re- rate summarized in Table 3, a weighted average of thecorrect classification rates based on the relative areassponse. The fact that even with this unrestricted nonpar-

ametric method, adjoining regions of ECa and crop yield of the soil types is taken for each soil.Given the ability of the statistical method to differen-are consistently categorized as one soil type, indicates

that a true relationship has been identified. tiate between the two major soil groups, we investigated

Table 4. Classification rates for individual soil types using electromagnetic induction and 1998–1999 no till full-season corn yields forBojac versus Wickham soil types.

1998 Corn 1999 Corn

Predicted soil type Predicted soil typeTrue soil Number of Number oftype observations Bojac Wickham observations Bojac Wickham

% %Bojac 74 93.2 6.8 105 98.1 1.9Wickham 55 10.9 89.1 92 5.4 94.6

1568 SOIL SCI. SOC. AM. J., VOL. 66, SEPTEMBER–OCTOBER 2002

Table 5. Percentage correct classification rate of classification trees for Bojac 1, Bojac 2, and Wickham (3 and 4) soil types.

Crop Number of observations ECa alone Crop yield alone ECa & crop yield

% correctly classifiedNo-till full-season corn 1998 129 72.1 78.3 85.3No-till full-season corn 1999 197 81.2 80.2 86.8No-till barley 1999 67 80.6 74.6 85.1No-till wheat 1999 211 73.9 81.0 91.5No-till double-crop soybean 1999 259 62.6 59.1 80.3

separating data further into variants within the Bojac into just Bojac and Wickham soils, we find that combin-(Bojac 1 and Bojac 2) and Wickham (Wickham 3 and ing ECa and crop yields improves the separation com-Wickham 4) soils. When attempting to classify the soil pared with the use of either of the variables consideredtypes into four groups, Bojac 1, 2 and Wickham 3 and alone (Table 5). Except for the soybean data, where an4, the correct classification rates dropped considerably, unusual precipitation pattern (170 mm of precipitationas the statistical method was unable to distinguish be- in 2 d in a September hurricane) resulted in few yieldtween the two Wickham soils effectively. This reflects differences between soils, the combined trees are ablethe similarity of the two variants of the Wickham soil. to correctly classify over 85% of the observations intoThe visually detectable differences between the two vari- the three soil types. Indeed, for the area of the fieldants of Wickham that lead to separate soil classification where the no-till double-crop soybeans were planted indo not translate into differences in water-holding capac- 1999, the ECa readings alone do not seem to be as ef-ity or yields. This result indicates the differences do not fective, as shown by the low correct classification ratetranslate to a broad difference in electrical conductivity. of 62.6% for ECa alone. The reason for this change inIn terms of soil mapping for determining soil maps for pattern is unknown. The 1999 wheat data was the bestvariable rate fertilization programs, the Wickham soil

classifier into the three soil groups with 91.5% beingvariants would receive the same fertilizer rate as they arecorrectly classified, while the corn and barley yieldsof the same soil productivity class (Simpson et al., 1993).combined with ECa readings gave correct classificationThus we tried separating the soils based on soil produc-rates approximately 85% (Table 5).tivity class using the ECa measurements to predict if the

Examining the results for the 1998 and 1999 no-tillsoil type was Bojac 1, Bojac 2, and the Wickham soilfull-season corn crops, we find similarities between clas-units combined. The productivity potential differencessification patterns for the three soil types. Figures 5 andbetween these three soil types are closely related to soil6 show the regions specified. The 1998 corn yield andwater-holding capacity (Roygard et al., 2002).ECa classification partition shows just three distinct re-Results from the different classification trees are sum-

marized in Table 5. As with the previous separation gions for the three soil types (Fig. 5). Given the nonpara-

Fig. 5. Classification partition for Bojac 1, 2 versus Wickham soil types using apparent electromagnectic conductivity (ECa) and 1998 no till full-season corn yields. The classification of each rectangle is determined by whichever soil type comprises the majority of observations.

ANDERSON-COOK ET AL.: DIFFERENTIATING SOIL USING CONDUCTIVITY AND YIELDS 1569

Fig. 6. Classification partition for Bojac 1, 2 versus Wickham soil types using apparent electromagnetic conductivity (ECa) and 1999 no till full-season corn yields. The classification of each rectangle is determined by whichever soil type comprises the majority of observations.

metric flexible model that has been used, this result is readings, high crop yields, or both. The actual observa-tions also show fewer clustering patterns than the 1998quite remarkable. Note how the actual observations

show a natural clustering into three groups. The un- data. For example, there seem to be high, medium, andlow values of corn yields for the Bojac 1 soil type. Whenshaded region indicates the Bojac 1 soil type, which is

characterized by low ECa readings and moderate corn we examine the misclassification of the soil types inTables 6 and 7 for 1998 and 1999 no-till full-season cornyields. The lightly shaded region indicates Bojac 2 soil,

which is made up primarily of low ECa readings and data, we see that the Wickham soil is most often cor-rectly classified, while some values of the Bojac soilslow corn crop yields. Finally, the Wickham soils are

shown on the partition with the darker shading, and are not easily separated. Clearly, it is more difficult todistinguish between the Bojac 1 and 2 soils than to sep-consist of high ECa readings, high corn yields or both.

The separation of soil types into these simple combina- arate them from the Wickham soils, as shown by thevaried misclassification rates in the different categories.tions of ECa readings and crop yields was also observed

in both the 1999 wheat and barley yields. There are many similarities in texture between the twoBojac soils (Table 1), which may explain their consistentThe 1999 corn crop classification has a similar correct

classification rate, but the regions for separating the ECa values.soils are not as easily characterized (Fig. 6). There aresome combinations of ECa readings and yields for the CONCLUSIONSBojac 1 and 2 soils, which are not easily distinguished.

Soil type classification using electromagnetic induc-This type of messy summary from a classification treetion with a statistical classification tree approach showsis not uncommon and was also observed for the 1999considerable promise for separating mid-Atlantic coastalsoybean data. However, the identification of the Wick-plain soils into broad ranges of potential productivity.ham soil continues to be the combination of high ECa

Table 7. Classification rates for individual soil types using appar-Table 6. Classification rates for individual soil types using ECa

and 1998 no till full-season corn yields for Bojac 1 and 2 versus ent electric conductivity ECa and 1999 corn yields for Bojac 1and 2 versus Wickham soil types.Wickham soil types.

Predicted soil type Predicted soil typeNumber of Number of

True soil type observations Bojac 1 Bojac 2 WickhamTrue soil type observations Bojac 1 Bojac 2 Wickham

% %Bojac 1 28 67.9 21.4 10.7 Bojac 1 24 45.8 50.0 4.2

Bojac 2 81 8.7 88.9 2.4Bojac 2 46 6.3 89.1 4.3Wickham 55 3.6 5.5 90.9 Wickham 92 0.0 4.3 95.7

1570 SOIL SCI. SOC. AM. J., VOL. 66, SEPTEMBER–OCTOBER 2002

Estimating depths to claypans using electromagnetic inductionFor separating soil types with significant subsoil texturemethods. J. Soil Water Conserv. 49:572–575.differences, such as Bojac and Wickham soils, the ECa Jaynes, D.B., T.S. Colvin, and J. Ambuel. 1995. Yield mapping by

readings alone are able to correctly classify the soil type electromagnetic induction. p. 383–394. In P. Robert et al. (ed.)Site-Specific Management for Agricultural Systems. Proceedingsover 85% of the time. When combined with a crop yieldof 2nd International Conf. 27–30 Mar. 1994. Minneapolis, MN.located with a GPS device, the correct classification rateASA, CSSA, and SSSA, Madison, WIincreases to over 90%. The ECa readings and the crop Kitchen, N.R., K.A. Sudduth, and S.T. Drummond. 1996. Mapping

yields are positively correlated, and high measurements of sand deposition from 1993 midwest floods with electromagneticinduction measurements. J. Soil Water Conserv. 51:336–340.of both ECa and crop yields are associated with the

McNeill, J.D. 1980a. Electrical Conductivity of soils and rocks. Techni-Wickham soils, while lower measurements of both arecal Note TN-5. Geonics Ltd., Mississauga, ON, Canada.associated with the Bojac soils. Depending on crop treat- McNeill, J.D. 1980b. Electromagnetic conductivity measurement at

ment, the rare combinations of high ECa and low crop low induction numbers. Technical Note TN-6, Geonics, Ltd., Mis-sissauga, ON, Canada.yield, or low ECa and high crop yields are categorized

McNeill, J.D. 1986. Geonics EM38 ground conductivity meter op-differently.erating instructions and survey interpretation techniques. Technical

Classifying soils into more precise soil types, such as Note TN-21. Geonics Ltd. Mississauga, ON, Canada.McNeill, J.D. 1992. Rapid, accurate mapping of soil salinity by electro-Bojac 1, 2, and Wickham soils is more difficult. Correct

magnetic ground conductivity meters. p. 209–229. In G.C. Topp etsoil classification rates using ECa readings alone rangeal. (ed.) Advances in measurements of soil physical properties:from 62 to 81%, and increases to between 80 and 91% Bringing theory into practice. SSSA Spec. Publ. 30. SSSA, Madi-

when combined with GPS crop yield information. Yield son, WI.Nicholson J., G. Hammer, D. Harper, and M. Crouch. 1998. Orderdata generally improve the classification over ECa data

1 Soil Survey of the cropping systems research site, Camden Farm,alone because crop yields tend to reflect soil moistureCaroline County. J.C. Nicholson, Soil Resource Specialist, Naturalholding capacities for coastal plain soils. Resources Conservation Service, Farmville, VA.

In this mid-Atlantic coastal plain study, electromag- Rhoades, J.D. 1981. Predicting bulk soil electrical conductivity vs.saturated paste extract electrical conductivity calibrations from soilnetic induction technology showed considerable prom-properties. Soil Sci. Soc. Am. J. 45:42–44.ise as a tool for soil classification, and could be a cost-

Rhoades, J.D., N.A. Manteghi, P.J. Shouse, and W.J. Alves. 1989.effective basis for development of variable rate fertilizer Soil electrical conductivity and soil salinity: New formulations and

calibrations. Soil Sci. Soc. Am. J. 53:433–439.application strategies.Robert, P. 1993. Characterization of soil conditions at the field level

for soil specific management. Geoderma 60:57–72.ACKNOWLEDGMENTS Roygard, J.K.F., M.M. Alley, and R. Khosla. 2002. No-till corn yields

and water balance in the mid-Atlantic coastal plain. Agron. J. 94:This research was supported by grants from the Foundation 612–622.

for Agronomic Research, the United Soybean Board, the Vir- Scanlon, B.R., J.G. Paine, and R.S. Goldsmith. 1999. Evaluation ofelectromagnetic induction as a reconnaissance technique to charac-ginia Corn Board, and the Virginia Small Grains Board. Appre-terize unsaturated flow in an arid setting. Ground Water. 37:ciation is expressed to John Nicholson, Greg Hammer, David296–304.Harper, and Mark Crouch (Natural Resource Conservation

Simpson, T.W., S.J. Donohue, G.W. Hawkins, M.M. Monnett, andService) for the Order I soil survey of the site and Steve J.C. Baker. 1993. The development and implementation of VirginiaDonohue and Steve Heckendorn for the soil sample analyses. agronomic land use evaluation system. Department of Crop and

Soil Environmental Science, Virginia Polytechnic Institute andState University, Blacksburg, VA.REFERENCES Soil Survey Staff. 1993. Soil systematics. p. 47–55. In Soil surveymanual, Nat. Res. Cons. Ser. USDA Handbook No. 18. USDA,Anderson-Cook, C.M., M.M. Alley, R. Noble, and R. Khosla. 1999.U.S. Gov. Print Off., Washington, DC.Phosphorous and potassium fertilizer recommendation variability

Sudduth, K.A., S.T. Drummond, and N.R. Kitchen. 2001. Accuracyfor two mid-atlantic coastal plains fields. Soil Sci. Soc. Am. J. 63:issues in electromagnetic induction sensing of soil electrical con-1740–1747.ductivity for precision agriculture. Comput. Electron. Agric. 31:Breiman, L., J.H. Friedman, R.A. Olshen, and C.J. Stone. 1984. Classi-239–264.fication and regression trees. Wadsworth and Brooks/Cole, Mon- Sudduth, K.A., N.R. Kitchen, and S.T. Drummond. 1999. Soil conduc-terey, CA. tivity sensing on claypan soils: comparison of electromagnetic in-

Clark, L.A., and D. Pregibon. 1991. Tree-based models. p. 377–420. duction and direct methods. p. 979–990. In P.C. Robert et al. (ed.)In J.M. Chambers and T.J. Hastie (ed.) Statistical models in S. Proc. 4th International Conference on Precision Agriculture. 19–22Wadsworth and Brooks/Cole, Pacific Grove, CA. July 1998, Minneapolis, MN. ASA, SSSA, and CSSA, Madison, WI.

Doolittle, J.A., E. Ealy, G. Secrist, D. Rector, and M. Crouch. 1995. S-Plus. 2000. User’s guide (statistical software), Mathsoft, Seattle,Reconnaissance soil mapping of a small watershed using electro- WA.magnetic induction and global positioning system techniques. Soil Zalasiewicz, J.A., S.J. Mathers, and J.D. Cornwall. 1985. The applica-Surv. Horiz. 36:86–94. tion of ground conductivity measurements to geological mapping.

Q. J. Engl. Geol. 18:139–148.Doolittle, J.A., K.A. Sudduth, N.R. Kitchen, and S.J. Indorante. 1994.